JP2012065281A - Communication program, communication apparatus, communication method, and communication system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To suppress that a protocol with communication performance lower than that of the other protocols is selected, when selecting a protocol.SOLUTION: A communication program causes a computer, which is included in a computer group and transmits data to the other computers included in the computer group by using any one of communication protocols among a plurality of communication protocols, to select a communication protocol, from the plurality of communication protocols, of which transfer time predicted on the basis of the number of hops on a communication path up to the other computers from the computer and a size of the data is shorter than that of the other communication protocols, and to transmit the data by using the selected communication protocol.

Description

  The present invention relates to a communication program, a communication device, a communication method, and a communication system. As a communication program, a communication device, and a communication method, for example, a communication program, a communication device, a communication method, and a communication system related to control of RDMA communication are included.

  A communication technique for transferring data stored in a memory in a computer to a memory in another computer is called RDMA (Remote Direct Memory Access). When data transfer is performed by RDMA, data is transferred to a designated transfer destination memory area in accordance with procedures such as the Eager protocol and the Rendezvous protocol.

  With regard to data transfer, for example, there is a technique for designating a message length of a method switching threshold (for example, see Non-Patent Document 1).

JP 2007-336101 A

“IBM System BlueGene Solution: Application Development”, June 2007, fifth edition, p. 5 “MVAPICH21.5 User Guide”, 11.21 MV2_IBA_EAGER_THRESHOLD, [online], July 11, 2010, [searched September 16, 2010], Internet (URL: http://mvapich.cse.ohio-state.edu /support/user_guide_mvapich2-1.5.html#x1-11000011.21)

  Since the number of communications required for data transfer differs depending on the protocol, the influence of the communication delay between the transfer destination and the transfer source on the time required for data transfer differs. For this reason, even when a protocol used for data transfer is selected according to the data size of the data to be transferred, depending on the communication delay, the time required for data transfer is better when another protocol not selected is used. May be less.

  An object of one aspect of the present invention is to suppress selection of a protocol having lower communication performance than other protocols.

  In order to solve the above problem, a communication program is a computer included in a computer group, and data is transmitted to another computer included in the computer group using any one of a plurality of communication protocols. The transmission time predicted based on the number of hops on the communication path from the computer to the other computer and the data size of the data among the plurality of communication protocols is transmitted to the transmitting computer. It is characterized by selecting a shorter communication protocol and transmitting the data using the selected communication protocol.

  According to one aspect of the present invention, when a protocol is selected, it is possible to suppress selection of a protocol having lower communication performance than other protocols.

It is a figure for demonstrating the eager protocol in this Embodiment. It is a figure for demonstrating the Rendezvous protocol in this Embodiment. It is a figure for demonstrating the number of hops. It is a figure which shows the example of a fat tree. It is a figure which shows the example of a connection of the computer group in the case of comprising a mesh. It is a figure which shows the example of a connection of the computer group in the case of comprising a torus. It is a figure which shows the structural example of the computer system in 1st embodiment. It is a figure which shows the hardware structural example of the computer in 1st embodiment. It is a figure which shows the function structural example of the computer in 1st embodiment. It is a figure which shows the structural example of a hop number management table. It is a flowchart for demonstrating an example of the process sequence of the transmission process of the data which the computer of a transmission side performs. It is a flowchart for demonstrating an example of the process sequence of the data reception process which the computer of the receiving side performs. It is a figure which shows the function structural example of the job management apparatus in 2nd embodiment. It is a figure which shows the structural example of a coordinate information storage part. It is a figure which shows the function structural example of the computer in 2nd embodiment. It is a sequence diagram for demonstrating an example of the process sequence at the time of the application execution start in 2nd embodiment.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. First, the concept of the present embodiment will be described.

  Transferring data stored in the main memory of one computer to the main memory of another remote computer is called RDMA (Remote Direct Memory Access). When transferring data d1 in the area r1 on the main memory of one computer c1 to the area r2 on the main memory of the other computer c2 using RDMA WRITE (data write request also called PUT), The computer c1 needs to know the virtual address of the area r2. In addition, data transfer is not always performed between the same areas. That is, transfer from the region r1 'to the region r2' and transfer from the region r1 "to the region r2" can also be performed. When data is transferred by RDMA, the eager protocol or the rendezvous protocol is used.

  FIG. 1 is a diagram for explaining the eager protocol in the present embodiment.

  In the figure, a computer C1 and a computer C2 each have an RDMA (Remote Direct Memory Access) communication function and are connected via a network such as an interconnect. A processing procedure for transferring data from the computer C1 to the computer C2 using the eager protocol will be described with reference to FIG.

  First, the computer C1 makes a memory copy of the transfer target data d1 stored in the area R1 in the main memory M1 to the transmission buffer R3 in the main memory M1 (S1). Subsequently, the computer C1 adds the control information h1 to either one before or after the data d1 in the transmission buffer R3 (S2). The control information h1 includes, for example, an eager protocol identifier. Subsequently, the computer C1 transfers the data d1 and the control information h1 in the transmission buffer R3 to the reception buffer R4 in the main memory M2 of the computer C2 (S3). The computer C1 knows in advance the virtual address of the reception buffer R4.

  Subsequently, the computer C2 decodes the control information h1 in the reception buffer R4 (S4). Subsequently, the computer C2 determines a storage area (area R2) and makes a memory copy of the data d1 to the area R2 in the main memory M2 (S5). In S5, the computer C2 can also perform processing such as saving the data d1 in a save area (not shown) before determining the storage area.

  In FIG. 1, regarding the transfer of data d1, the time (T_Eger) required for communication in the eager protocol is represented by the following equation (E), for example.

T_Eager = (D / W) + (L_1 + D / B) + (D / W) + S_Eager (E)
D: Data size (bytes)
W: Memory copy bandwidth (bandwidth) (byte / sec)
B: Interconnect bandwidth (RDMA communication bandwidth) (byte / sec)
L_1: RDMA communication delay of the interconnect (sec)
S_Eger: Software overhead time in the eager protocol (sec)

  Here, (D / W) in the first term on the right side of the equation (E) is the time required for the memory copy in step S1. The second term (L_1 + D / B) is the time required for data transfer in step S3. The third term (D / W) is the time required for the memory copy in step S5. S_Eager in the fourth term is an overhead time required for software processing including steps S2 and S4.

  Note that the communication delay of L_1 means an overhead time by hardware required for transferring 0-byte data.

  On the other hand, FIG. 2 is a diagram for explaining the Rendezvous protocol in the present embodiment. The relationship between the computer C1 and the computer C2 is the same as in FIG. A processing procedure for transferring data from the computer C1 to the computer C2 using the Rendezvous protocol will be described with reference to FIG.

  First, the computer C1 transmits control information h2 to the computer C2 (S11). The control information h2 includes, for example, an Rendezvous protocol identifier. When the computer C2 receives the control information h2, the computer C2 decodes the control information h2 (S12). Subsequently, the computer C2 transmits the control information h3 to the computer C1 (S13). The control information h3 includes, for example, a virtual address of the reception area R2 in the main memory M2.

  When the computer C1 receives the control information h3, the computer C1 decodes the control information h3 and obtains a virtual address or the like of the region R2 (S14). Subsequently, the computer C1 transfers the data d1 stored in the area R1 in the main memory M1 to the area R2 of the computer C2 (S15).

  In FIG. 2, the time (T_Rendezvous) required for communication in the Rendezvous protocol regarding the transfer of the data d1 is represented by the following equation (R), for example.

T_Rendezvous = L_2 × 2 + (L_1 + D / B) + S_Rendezvous (R)
D: Data size (bytes)
B: Interconnect bandwidth (bandwidth) (byte / sec)
L_1: RDMA communication delay of the interconnect (sec)
L_2: Communication delay of interconnect control communication (sec)
S_Rendezvous: Software overhead time in the Rendezvous protocol (sec)

  Here, the first term L_2 × 2 on the right side of the equation (R) is the time required to transmit / receive the control information h2 or h3 in steps S11 and S13. Since the control information h2 and h3 have a very small data size, it can be considered that only the communication delay L_2 is required. The second term (L_2 + D / B) is the time required for step S15. The third term S_Rendezvous is an overhead time required for software processing including steps S12 and S14.

  Regarding the equations (E) and (R), if the network topology has a substantially uniform communication time between arbitrary nodes (computers), in general, W, B, L_1, L_2, S_Eager, and S_Eager The value of depends on the characteristics of the hardware and software. Therefore, the values of these parameters can be considered to be almost constant regardless of the combination of computers that communicate.

  In a normal network state, the smaller the data size D, the more T_Eager << T_Rendezvous. On the other hand, T_Eager >> T_Rendezvous as the data size D increases. This is apparent when comparing the case where 0 is substituted for D and the case where ∞ (infinity) is substituted for D in the equations of T_Eager and T_Rendezvous.

  In addition, the expressions (E) and (R) are both linear expressions of the data size D. Therefore, the threshold value D_Threshold of the data size D for switching between the Eager protocol and the Rendezvous protocol can be obtained by solving D in the case of T_Eager = T_Rendezvous with respect to the expressions (E) and (R). The result is the following equation (Dt1).

  D_Threshold = (L_2 + (S_Rendezvous−S_Eager) / 2) × W (Dt1)

  When the communication time between arbitrary nodes in the network topology is substantially uniform, for example, data of data to be transferred is more than D_Threshold obtained by substituting values of W, L_2, S_Eager, and S_Eager as constants. By switching the protocol based on whether the size is large or small, the time required for communication is suppressed.

  FIG. 3 is a diagram for explaining the number of hops. Note that the hop number H refers to the number of connections through the interconnect on the communication path between the two computers Ci. The nodes N1 to N5 are computers or switches. For example, when the node N1 is the transmission source and the node N5 is the transmission destination, the hop count H is 4.

  Examples of the network topology include a fat tree, a mesh, and a torus.

  FIG. 4 is a diagram illustrating an example of a fat tree. In the fat tree shown in the figure, the root switches SW1 and SW2 are switches as root nodes. Leaf switches SW3 to SW6 are connected to the route switches SW1 and SW2, respectively. Computers c31 to c34, c41 to c44, c51 to c54, or c61 to c64 as calculation nodes are connected to the leaf switches SW3 to SW6, respectively.

  In a normal tree network topology, since there is only one root node, the communication load is concentrated in the vicinity of the root node, which causes a decrease in communication performance. In order to avoid the disadvantages of the normal tree, a plurality of root nodes are arranged in the fat tree network topology.

  In the fat tree, the values of W, B, S_Eager, and S_Eager can be considered to be substantially constant. For L_1 and L_2, when communicating via a root node (for example, communication between c31 and c64 or communication between c31 and c51) and when communicating without passing through a root node (for example, communication between c31 and c32) Since the number of hops H is different, the values are expected to be different. However, most communication nodes whose communication destinations are included in the fat tree (c41 to c44, c51 to c54, and c61 to c64 for c31) are expected to have substantially the same value.

  Therefore, when the computers C1 and C2 shown in FIG. 1 or FIG. 2 are arranged as calculation nodes in the fat tree, by switching between the Eager protocol and the Rendezvous protocol according to the data size D of the data d1, etc. It is considered that the communication time between the computer C1 and the computer C2 is shortened.

  FIG. 5 is a diagram illustrating a connection example in which the computers C1 to C9 form a mesh. FIG. 6 is a diagram illustrating a connection example in which the computers C1 to C9 constitute a torus. A line segment connecting each computer Ci indicates an interconnect connecting the computers Ci related to both ends of the line segment.

  In FIG. 5, a two-dimensional mesh is shown. In FIG. 6, a two-dimensional torus is shown. However, the number of dimensions is not limited to a predetermined number.

  As shown in FIGS. 5 and 6, the number of hops between arbitrary nodes is not determined in the mesh or the torus.

  Depending on the network topology, the values of L_1 and L_2 change depending on the combination of computers that perform communication. L_1 or L_2 is expressed as follows as a linear function of the number of hops H between computers performing communication.

L_1 = L_1N + A_1 × H (L1)
L_2 = L_2N + A_2 × H (L2)

  Here, A_1 and A_2 are the amount of increase (sec / hop) in interconnect communication delay (communication delay of RDMA communication) per hop, respectively. Normally, the values of A_1 and A_2 match, but different variables are assigned to the formula (L1) and the formula (L2) for generalization.

  L_1N and L_2N are hardware overhead times (sec) in the communication delay of the first hop (between the transmission source and the adjacent node).

  The values of A_1, A_2, L_1N, and L_2N may be measured in advance in a computer system that is actually used. Specifically, for A_1, it is only necessary to measure the amount of increase in communication delay every time the number of hops increases by one during communication when the eager protocol is used. L_1N may be obtained by subtracting A_1 × H from the actual communication delay during communication with the number of hops H when the eager protocol is used.

  Similarly, for A_2, it is only necessary to measure the amount of increase in communication delay every time the number of hops increases by one during communication when the Rendezvous protocol is used. L_2N may be obtained by subtracting A_2 × H from the actual communication delay during communication with the number of hops H when the Rendezvous protocol is used.

  The fact that L_1 and L_2 are affected by the hop number H means that the threshold D_Threshold is also affected by the hop number H. Therefore, for a network topology that cannot handle the values of L_1 and L_2 as constants, the threshold D_Threshold calculated by the equation (Dt1) is insufficient. That is, it is considered that there is room for further improving the communication performance by devising a method for calculating the threshold value D_Threshold.

  Therefore, the following formulas (Eh) and (Rh) can be obtained by substituting the formulas (L1) and (L2) into the formulas (E) and (R).

  T_Eager = (D / W) × 2 + (L — 1N + A — 1 × H) + D / B + S_Eager (Eh)

  T_Rendezvous = (L_2N + A_2 × H) × 2 + (L_1N + A_1 × H) + D / B + S_Rendezvous (Rh)

  That is, T_Eager and T_Rendezvous are linear expressions of the hop count H.

  The threshold D_Threshold for switching between the eager protocol and the rendezvous protocol may be obtained by solving the expressions (Eh) and (Rh) for D when T_Eger = T_Rendezvous.

That is,
(2 / W) × D = (L_2N + A_2 × H) × 2 + (S_Rendezvous−S_Eager)
From
D_Threshold = [A_2 × H + L_2N + (S_Rendezvous−S_Eager) / 2] × W (Dt2)
Is guided.

  In the present embodiment, the communication protocol to be used is switched according to the comparison between the calculation result of the formula (Dt2) and the data size of the data D. Formula (Dt2) is a formula derived in consideration of the fact that L_1 and L_2 change according to the number of hops. Therefore, according to the threshold value D_Threshold based on the equation (Dt2), it is possible to select a communication protocol to be used in consideration of not only the data size of the data D but also the number of hops H between computers performing communication. As a result, compared with the case where a communication protocol is selected based on the formula (Dt1), improvement in communication performance can be expected particularly in a mesh or torus network topology.

  An example of a specific computer system to which the above concept is applied will be described.

  FIG. 7 is a diagram illustrating a configuration example of the computer system according to the first embodiment. In the figure, a job management apparatus 20 and a computer group Cs are connected via a network 30 such as a LAN (Local Area Network).

  The job management apparatus 20 is a computer that receives a job input from a user, determines an assignment order (dispatch order) of the input job to the computer group Cs, and dispatches the job. The acceptance of the job input may be input by directly operating the job management apparatus 20 or may be input by communication with the terminal apparatus 50 via the network 40. Note that the job management apparatus 20 is an example of an information processing apparatus.

  The computer group Cs is a set of distributed memory type parallel computers. Each computer Ci (hereinafter, i is assigned a different integer for each computer) is connected by a line (network) such as an interconnect and has an RDMA communication function.

  For example, the computer group Cs is connected to configure a mesh or a torus as a network topology.

  FIG. 8 is a diagram illustrating a hardware configuration example of the computer according to the first embodiment. In the figure, the computer Ci includes a drive device 100, an auxiliary storage device 102, a main storage device 103, a CPU 104, and an interface device 105, which are mutually connected by a bus B.

  A program for realizing processing in the computer Ci is provided by a recording medium 101 such as a CD-ROM. When the recording medium 101 on which the program is recorded is set in the drive device 100, the program is installed from the recording medium 101 to the auxiliary storage device 102 via the drive device 100. However, the program need not be installed from the recording medium 101 and may be downloaded from another computer via a network. The auxiliary storage device 102 stores the installed program and also stores necessary files and data.

  The main storage device 103 reads the program from the auxiliary storage device 102 and stores it when there is an instruction to start the program. The CPU 104 executes a function related to the computer Ci according to a program stored in the main storage device 103. The interface device 105 is used as an interface for connecting to a network (interconnect).

  FIG. 9 is a diagram illustrating a functional configuration example of the computer according to the first embodiment. In the figure, the computer C1 is the data transmission side, and the computer C2 is the data reception side.

  The computer C1 includes an application 11, a transmission control unit 12, and the like. The application 11 is a program that executes predetermined processing using RDMA communication. Each application 11 is started as a process in the computer Ci to which a job is assigned by the job management apparatus 20, for example.

  The transmission control unit 12 controls data transmission by RDMA communication in response to a data transmission request from the application 11. The transmission control unit 12 is realized by a process that a program installed in the computer C1 causes the CPU 104 of the computer C1 to execute. The transmission control unit 12 is implemented as a part of an MPI (Message Passing Interface) library, for example.

  In the figure, the transmission control unit 12 includes a threshold value calculation unit 121, a parameter storage unit 122, a protocol selection unit 123, an E transmission control unit 124, an R transmission control unit 125, and the like.

  The threshold value calculation unit 121 calculates the threshold value D_Threshold based on the above formula (Dt2). The parameter storage unit 122 stores various parameters necessary for calculating the threshold value D_Threshold using the auxiliary storage device 102, for example. Specifically, the parameter storage unit 122 stores values or theoretical values measured in advance for W, L_2N, A_2, S_Rendezvous, and S_Eager.

  The parameter storage unit 122 also stores a hop number management table 122t in which the hop number H of the shortest communication path from the computer Ci to another computer Cj is registered.

  FIG. 10 is a diagram illustrating a configuration example of the hop number management table. As shown in the figure, the hop count management table 122t stores the hop count H for each computer number. The hop number H is a relative value starting from the computer Ci storing the hop number management table 122t. The figure shows an example in which the number of hops H from the computer C1 to each computer Cj when the mesh shown in FIG. 5 is configured is stored. The computer number is a number for identifying each computer Ci (strictly speaking, the process of the application 11 operating on each computer).

  The protocol selection unit 123 selects a communication protocol to be used based on a comparison between the data size of the data requested to be transmitted from the application 11 and the threshold value D_Threshold calculated by the threshold value calculation unit 121. The E transmission control unit 124 controls data transmission processing by the eager protocol. The R transmission control unit 125 controls data transmission processing according to the Rendezvous protocol.

  On the other hand, the computer C2 on the receiving side includes the application 11, the reception control unit 13, and the like. The application 11 is as described above. However, on the receiving side, the application 11 requests the reception control unit 13 to receive data.

  The reception control unit 13 controls reception of data by RDMA communication in response to a data reception request from the application 11. The reception control unit is realized by processing that the program installed in the computer C2 causes the CPU 104 of the computer C1 to execute. The reception control unit 13 is implemented as a part of the MPI library, for example.

  In the figure, the reception control unit 13 includes a distribution unit 131, an E reception control unit 132, an R reception control unit 133, and the like. The distribution unit 131 determines which communication protocol is selected on the transmission side, and distributes the execution subject of the reception process to the E reception control unit 132 or the R reception control unit 133. The E reception control unit 132 controls data reception processing by the eager protocol. The R reception control unit 133 controls data reception processing according to the Rendezvous protocol.

  Note that the transmission side and the reception side have a relative relationship. That is, each computer Ci is a transmitting side at a certain time and a receiving side at a certain time. Therefore, each computer Ci has both the transmission control unit 12 and the reception control unit 13. The application 11 of each computer Ci transmits data using the transmission control unit 12 and receives data using the reception control unit 13.

  Hereinafter, a processing procedure executed by the computer Ci will be described. FIG. 11 is a flowchart for explaining an example of a processing procedure of data transmission processing executed by the transmission-side computer.

  In step S <b> 101, the transmission control unit 12 receives a transmission request for data d <b> 1 from the application 11. In the transmission request, parameters such as data d1, data size D of the data d1, and a transmission destination computer number are designated.

  Subsequently, the threshold value calculation unit 121 acquires W, L_2N, A_2, S_Rendezvous, S_Eager, and the hop count H, which are parameters for calculating the threshold value D_Threshold, from the parameter storage unit 122 (S102). As the hop number H, a value associated with the computer number of the transmission destination computer Cj is acquired from the hop number management table 122t. Subsequently, the threshold value calculation unit 121 calculates the threshold value D_Threshold by substituting the acquired parameter into the equation (Dt2) (S103).

  Subsequently, the protocol selection unit 123 selects the communication protocol to be used by comparing the data size D with the threshold value D_Threshold (S104). When the data size D is smaller than the threshold value D_Threshold (Yes in S104), the protocol selection unit 123 selects the eager protocol. In response to the selection, the E transmission control unit 124 executes the transmission process of the data d1 via the interconnect according to the processing procedure described in steps S1 to S3 in FIG. 1 (S105).

  On the other hand, when the data size D is equal to or greater than the threshold value D_Threshold (No in S104), the protocol selection unit 123 selects the Rendezvous protocol. In response to the selection, the R transmission control unit 125 performs transmission of data d1 through the interconnect in the procedure described in steps S11 to S15 in FIG. 2 (S106).

  In FIG. 11, when the data size D matches the threshold value D_Threshold, the Rendezvous protocol is selected, but the eager protocol may be selected.

  Next, FIG. 12 is a flowchart for explaining an example of a processing procedure of data reception processing executed by the receiving computer.

  In step S201, the reception control unit 13 receives a reception request for data d1. Subsequently, the distribution unit 131 waits for information to be received via the interconnect (S202). When the information is received (Yes in S202), the distribution unit 131 determines which communication protocol is selected on the transmission side by decoding the received information (S203). That is, when the eager protocol is selected on the transmission side, the information received first is the control information h1 and the data d1 (see FIG. 1). On the other hand, in the case of the Rendezvous protocol, the information received first is the control information h2 (see FIG. 2). Therefore, the distribution unit 131 decodes the control information h1 or the control information h2, and determines the communication protocol selected on the transmission side.

  When determining that the eager protocol has been selected on the transmission side (Eager in S204), the distribution unit 131 delegates subsequent reception processing to the E reception control unit 132. Therefore, the E reception control unit 132 executes the reception process for the data d1 via the interconnect in accordance with the processing procedure described in steps S3 to S5 in FIG. 1 (S205).

  On the other hand, when determining that the Rendezvous protocol is selected on the transmission side (Rendezvous in S204), the distribution unit 131 delegates subsequent reception processing to the R reception control unit 133. Therefore, the R reception control unit 133 executes the reception process of the data d1 via the interconnect in the processing procedure described in steps S11 to S15 in FIG. 2 (S206).

  As described above, according to the first embodiment, the threshold D_Threshold for determining the communication protocol to be used is determined in consideration of the number of hops. Here, the threshold value D_Threshold may differ depending on the computer Ci that is the communication partner. Therefore, even when a network topology such as a mesh or a torus where communication time between arbitrary nodes is not uniform is configured, a more effective communication protocol can be selected from the viewpoint of communication performance.

  Note that this embodiment is applicable even when the network topology is a fat tree. In order to explain this point, equations (Dt1) and (Dt2) are shown below again.

  D_Threshold = (L_2 + (S_Rendezvous−S_Eager) / 2) × W (Dt1)

  D_Threshold = [A_2 × H + L_2N + (S_Rendezvous−S_Eager) / 2] × W (Dt2)

  As apparent from the above, the formula (Dt1) and the formula (Dt2) are different from each other in that L_2 in the formula (Dt1) is replaced with A_2 × H + L_2N in the formula (Dt2). When the amount of change in the number of hops between arbitrary nodes is small, for example, a fixed value (for example, 0) may be substituted for the number of hops H in Expression (Dt2). In that case, the expression (Dt2) is an approximate expression of the expression (Dt1).

  Further, the expression (Dt2) may be simplified within a range that is permitted in operation. For example, when the software overhead time in the Eager protocol and the software overbed time in the Rendezvous protocol are negligible in terms of operation, the term (S_Rendezvous−S_Eager) / 2 is removed from the equation (Dt2), (Dt3) may be used.

D_Threshold = (A_2 × H + L_2N) × W (Dt3)
Alternatively, when the value of L_2N is a value that can be ignored in operation, the following equation (Dt4) in which the term of L_2N is removed from the equation (Dt2) may be used.

  D_Threshold = [A_2 × H + (S_Rendezvous−S_Eager) / 2] × W (Dt4)

  Furthermore, the following equation (Dt5) in which both (S_Rendezvous−S_Eager) / 2 and L_2N are removed may be used.

  D_Threshold = (A_2 × H) × W (Dt5)

  In other words, when the formula (Dt4) or the formula (Dt2) is used, the threshold D_Threshold can be calculated in consideration of the software overhead time. Further, when the formula (Dt3) or the formula (Dt2) is used, the threshold value D_Threshold can be calculated in consideration of the value of L_2N.

  By the way, in the first embodiment, the hop number management table 122t is stored in each computer Ci. The amount of information in the hop number management table 122t increases as the number of computers Ci increases. The contents of the hop number management table 122t are different for each computer Ci. Therefore, when a large number of computers Ci are used, the burden of setting the hop number management table 122t becomes very large. Also, when the connection relationship of the computers Ci changes or when the number of computers Ci increases or decreases, it is very difficult to update the hop number management table 122t of each computer Ci according to the new configuration. is there.

  Therefore, in the second embodiment, an example in which the maintenance work of the hop number management table 122t is simplified will be described. In the second embodiment, differences from the first embodiment will be described. Accordingly, the points not particularly mentioned may be the same as those in the first embodiment.

  FIG. 13 is a diagram illustrating a functional configuration example of the job management apparatus according to the second embodiment. In FIG. 1, the job management apparatus 20 includes a job allocation unit 21, a hop number calculation unit 22, a coordinate information storage unit 23, and the like. Any computer Ci of the computers Cs may include the job allocation unit 21, the hop number calculation unit 22, and the coordinate information storage unit 23 to calculate the hop number. Alternatively, the computer Ci of the computers Cs may include the hop number calculation unit 22 and calculate the hop number based on the coordinate values stored in the coordinate information storage unit 23 included in the job management apparatus.

  The job allocation unit 21 receives an instruction to execute the application 11 from the user, and instructs the computer Cn to execute the application 11 (that is, execute the job).

  The coordinate information storage unit 23 stores the coordinate value (position information) of each computer Ci in the network topology coordinate system using, for example, an auxiliary storage device of the job management device 20.

  That is, in the network topology of mesh or torus, the computer Ci is usually arranged in the form of an n-dimensional rectangular parallelepiped on logical lattice points in the n-dimensional coordinate space. If the n-dimensional coordinate system is expressed as (x0, x1,..., xn-1), the computer Ci is converted into an n-dimensional cuboid xi_min <= xi <= xi_max (i = 0,..., n-1). Be placed. Accordingly, each computer Ci has coordinate values (x0, x1,..., Xn−1). The coordinate information storage unit 23 stores such coordinate values for each computer Ci.

  FIG. 14 is a diagram illustrating a configuration example of the coordinate information storage unit. As shown in the figure, the coordinate information storage unit 23 stores coordinate values in the network topology coordinate system for each computer number. In the figure, an example is shown in which the coordinate values of each computer Ci are stored when the network topology is the two-dimensional mesh shown in FIG.

  The hop count calculation unit 22 calculates the hop count H between the computers Ci based on the coordinate values recorded in the coordinate information storage unit 23. The calculation method of the hop number H differs depending on the network topology.

  A method for calculating the hop collection will be described. In the mesh network topology, the computer Ci located at xi_min and the computer Ci located at xi_max are not directly connected by an interconnect. For example, in FIG. 5, the computer C1 (x1_min, x2_min) and the computer C3 (x1_max, x2_min) are not directly connected. Therefore, when the coordinates of the computer C1 in the n-dimensional rectangular parallelepiped are (x0_1, x1_1, ..., xn-1_1) and the coordinates of C2 are (x0_2, x1_2, ..., xn-1_2), the shortest interval between them The number of hops H of the communication path is calculated by the following equation (Hm).

  Σ | xi_1-xi_2 | = | x0_1-x0_2 | + | x1_1-x1_2 | +. . . + | Xn-1_1-xn-1_2 | ... (Hm)

  For example, the hop number H of the shortest communication path between adjacent computers is 1.

  On the other hand, in the torus network topology, the computer Ci located at xi_min and the computer Ci located at xi_max are directly connected by an interconnect. For example, in FIG. 6, the computer C1 (x1_min, x2_min) and the computer C3 (x1_max, x2_min) are directly connected. Therefore, the calculation formula of the hop count H of the shortest communication path between the computer C1 and the computer C2 becomes more complicated.

The length Diameter_i of the i-dimensional side of the n-dimensional rectangular parallelepiped of the torus network topology is
Diameter_i = xi_max−xi_min + 1
And

  Subsequently, a length Radius_i that is ½ of Diameter_i is obtained as follows.

  Radius_i = (xi_max-xi_min + 1) / 2

  The coordinates of the computer C1 in the n-dimensional rectangular parallelepiped are (x0_1, x1_1, ..., xn-1_1), and the coordinates of C2 are (x0_2, x1_2, ..., xn-1_2). In this case, the hop count H of the shortest communication path between the computer C1 and the computer C2 is calculated by the following formula (Ht1) or formula (Ht2).

If | xi_1-xi_2 | <= Radius_i,
Σ | xi_1-xi_2 | (Ht1)
If | xi_1-xi_2 |> Radius_i,
Σ (Diameter_i- | xi_1-xi_2 |) (Ht2)

  From the above, if the network topology is a mesh, the hop number calculation unit 22 calculates the hop number H based on the formula (Hm). Further, if the network topology is a torus, the hop number calculation unit 22 calculates the hop number H based on the formula (Ht1) or the formula (Ht2).

  FIG. 15 is a diagram illustrating a functional configuration example of a computer according to the second embodiment. In FIG. 15, the same parts as those of FIG. 9 are denoted by the same reference numerals, and the description thereof is omitted.

  In FIG. 15, the initialization unit 14 is clearly shown in the computer Ci. The initialization unit 14 performs an initialization process that is required before the execution of the communication process. The initialization unit 14 acquires the hop count H and the like from the job management apparatus 20 in the initialization process. That is, the initialization unit is an example of an acquisition unit. The initialization unit 14 is implemented as a part of the MPI library, for example. In this case, the initialization unit 14 corresponds to a function for initialization.

  Hereinafter, a processing procedure of the computer system according to the second embodiment will be described. FIG. 16 is a sequence diagram for explaining an example of a processing procedure at the start of application execution in the second embodiment.

  In step S301, the job assignment unit 21 of the job management apparatus 20 receives an execution instruction for the application 11 from the user. In the execution instruction, the number of computers Ci to be used is also specified.

  Subsequently, the job allocation unit 21 selects, from the computer group Cs, as many computers Ci designated by the user as the execution destination of the application 11 (S302). Specifically, the computer number of the computer Ci used as the execution destination of the application 11 is determined. The computer Ci may be selected based on a known job scheduling technique or the like. Subsequently, the job assignment unit 21 transmits an execution instruction of the application 11 to each selected computer Ci (S303).

  Each computer Ci instructed to execute the application 11 activates the application 11 (S304). In response to activation, the application 11 requests the initialization unit 14 to execute initialization processing (S305). The initialization unit 14 inquires about the computer numbers and the hop numbers H of all the computers Ci selected as the execution destination of the application 11 during the initialization process (S306). In the inquiry, the computer number of the computer Ci as the inquiry source is also specified.

  The hop number calculation unit 22 of the job management apparatus 20 acquires the computer number of the inquiry source computer Ci and the coordinate value of the other computer Ci selected as the execution destination of the application 11 from the coordinate information storage unit 23 (S307). ). Subsequently, the hop count calculation unit 22 calculates the hop count H from the inquiry computer Ci to the other computer Ci based on the acquired coordinate value (S308). That is, the hop count H is calculated based on the positional relationship between the computer Ci that is the inquiry source and the other computer Ci. Note that since the network topology is statically determined, it is determined in advance which one of the formula (Hm), the formula (Ht1), and the formula (Ht2) is used.

  Subsequently, the hop number calculation unit 22 returns the computer number of the other computer Ci and the hop number H to the computer Ci that has made the inquiry (S309). When there are a plurality of other computers Ci, a plurality of sets of computer numbers and hop numbers H are returned.

  Subsequently, the initialization unit 14 records the received computer number and hop number H in the hop number management table 122t of the parameter storage unit 122 (S310).

  The communication process executed after that may be the same as in the first embodiment.

  As described above, according to the second embodiment, the hop count H of another computer Ci is automatically registered for each computer Ci. Therefore, it is possible to remarkably reduce the work burden of the registration work of the hop number H to each computer Ci. That is, the administrator or the like may edit the coordinate information storage unit 23 that is centrally managed in the job management apparatus 20.

  Note that the hop count management table 122t distributed to each computer Ci in the first embodiment may be centrally stored in the job management apparatus 20 instead of the coordinate information storage unit 23. Also in this case, the hop count H can be automatically registered in each computer Ci by the same processing procedure as in FIG. In this case, the hop number calculation unit 22 does not need to calculate the hop number H. The hop count calculation unit 22 may return the hop count H or the like based on the hop count management table 122t stored with respect to the inquiry source computer Ci.

  However, the hop number management table 122t needs to be created for the number of computers Ci. Therefore, it is considered that the work of registration in the coordinate information storage unit 23 is smaller.

  As mentioned above, although the Example of this invention was explained in full detail, this invention is not limited to such specific embodiment, In the range of the summary of this invention described in the claim, various deformation | transformation・ Change is possible.

11 Application 12 Transmission Control Unit 13 Reception Control Unit 14 Initialization Unit 20 Job Management Device 21 Job Allocation Unit 22 Hop Number Calculation Unit 23 Coordinate Information Storage Unit 100 Drive Device 101 Recording Medium 102 Auxiliary Storage Device 103 Main Storage Device 104 CPU
105 interface device 121 threshold calculation unit 122 parameter storage unit 123 protocol selection unit 124 E transmission control unit 125 R transmission control unit 131 distribution unit 132 E reception control unit 133 R reception control unit B bus Cs computer group Ci computer

Claims (7)

A computer included in a computer group, the computer transmitting data to another computer included in the computer group using any one of a plurality of communication protocols,
Of the plurality of communication protocols, a communication protocol in which the transfer time predicted based on the number of hops on the communication path from the computer to the other computer and the data size of the data is shorter than other communication protocols Select
A communication program that causes the data to be transmitted using a selected communication protocol. In the computer,
Based on the information indicating the connection relationship between the computers included in the computer group stored in the storage means in advance, calculate the number of hops on the communication path from the computer to the other computer,
The communication protocol is selected from among the plurality of communication protocols, wherein a communication protocol predicted based on the calculated number of hops and a data size of the data is shorter than other communication protocols. Item 4. The communication program according to Item 1. The plurality of communication protocols include communication protocols in which the number of times of communication between the computer and the other computer required for the computer to transmit the data to the other computer is different from each other. Item 3. A communication program according to item 1 or 2. In the computer,
Based on the number of hops, a threshold for switching a communication protocol having a shorter transfer time than other communication protocols included in the plurality of communication protocols is set for the data size,
4. The method according to claim 1, further comprising: selecting a communication protocol having a smaller transfer time than other communication protocols according to the set threshold value and the data size of the data. 5. The communication program described in the section. A communication device that transmits data to other communication devices included in the communication device group using any one of a plurality of communication protocols,
Of the plurality of communication protocols, the transfer time predicted based on the number of hops on the communication path from the communication device to the other communication device and the data size of the data is shorter than other communication protocols. A selection means for selecting a communication protocol;
And a transmission control means for transmitting the data using the selected communication protocol. A computer included in the computer group, the computer transmitting data to another computer included in the computer group using any one of a plurality of communication protocols,
Of the plurality of communication protocols, a communication protocol in which the transfer time predicted based on the number of hops on the communication path from the computer to the other computer and the data size of the data is shorter than other communication protocols Select
A communication method, comprising: transmitting the data using a selected communication protocol. Communication having a communication device that transmits data to another communication device included in the communication device group using any one of a plurality of communication protocols, and an information processing device not included in the communication device group A system,
Obtaining means for obtaining from the information processing device the number of hops on the communication path from the communication device to the other communication device;
A selection unit that selects a communication protocol whose transfer time predicted based on the number of hops acquired by the acquisition unit and the data size of the data is shorter than other communication protocols among the plurality of communication protocols means,
Transmission control means for transmitting the data using a selected communication protocol,
The information processing apparatus includes:
The number of hops on a communication path from the communication device to the other communication device is calculated based on information indicating a connection relationship between the communication devices included in the communication device group stored in the storage unit in advance. A communication system.

Images